Desktop AI editor for Windows, macOS (Apple Silicon), and Linux
The AI editor that does not mark up your tokens.
$20 a month is the editor seat. Chat leaves your machine only for Anthropic, xAI, or OpenAI on your own keys, billed at their rates. There is no Pumpkin-hosted server in that path, so Pumpkin never receives your prompt, and there is no extra network hop to wait on.
Built on Code-OSS. Your Cursor rules and MCP config still work.
- 01No upcharge on tokensThe seat pays for the editor. Inference is billed by the provider.
- 02Bring your own modelsAnthropic, xAI, or OpenAI keys in the OS keychain.
- 03No Pumpkin-hosted serverNo Pumpkin-hosted server sits between your machine and the provider.
- 04The full agent toolkitSubagents, rules, MCP, browser, canvas, on your machine.
01
No upcharge on tokens
$20 a month buys the editor. That is the whole bill from us. Chat tokens are metered by Anthropic, xAI, or OpenAI at the rates on their own price pages, on your own account, with nothing added.
Pumpkin does not resell inference. There is no hosted usage pool, no Pumpkin request quota, and no "fast" tier to buy back your speed.
02
Bring your own models
Paste an Anthropic key, an xAI key, an OpenAI key, or all three into Pumpkin settings. The keys go into the operating system keychain, not a Pumpkin config file and not our servers. Pick the provider per chat from the model chip in the agent pane.
Chat is Anthropic, xAI, and OpenAI, including OpenAI's gpt-5.6 models. Semantic code search is separate: grepai embeds with a local model, or with an OpenAI embedding key if you add one. That embedding key is a different credential from the OpenAI chat key and is used only for search embeddings.
03
No Pumpkin-hosted server
When you send a message, chat may go through a loopback process on the same machine. It leaves your computer only for Anthropic, xAI, or OpenAI, whichever provider you picked for that chat. There is no Pumpkin-hosted server, no remote relay, and no extra network hop. The agent loop, file edits, terminal, ripgrep and grepai search, and MCP client connections and tools are managed from the machine in front of you. Configured MCP servers may be remote.
Two things follow. Privacy: Pumpkin never receives the prompt, the code, or the reply. Speed: nothing Pumpkin-hosted sits in the chat path and there is no remote queue to wait in. Your Pumpkin account on pumpkin.club handles sign-in and the license, not chat.
04
The full agent toolkit, on your machine
Everything you expect from a modern agent editor, running locally. Change the prompt and resend, and the files go back to where they were before that turn.
- Subagents Spawn, watch, and message child agents on the same worktree.
- Rules and skills AGENTS.md, rules, skills, and slash commands from your workspace and home folder.
- MCP mcp.json from
.pumpkinand.cursoris merged; big tool lists defer to list-and-call. - Browser Real in-window tabs the agent drives; take one over any time.
- Canvas Live React or HTML built beside the chat.
- grepai and WSL Local semantic search;
pumpkin .in WSL opens the Windows UI on a Linux remote.
Why switch?
Because your editor should not be a middleman for your models.
- $20 editor seatOne flat price for the whole editor. That is the entire bill from Pumpkin.
- Provider-rate tokensAnthropic, xAI, and OpenAI bill you at the rates on their own price pages, on your own account. Pumpkin adds nothing.
- No remote relayYour prompt leaves your machine only for the provider. No Pumpkin-hosted server, no remote relay, nobody else reading along.
- Full desktop toolkitAgent, subagents, rules, MCP, browser, canvas, terminal, and search, all on your machine, all included in the seat.
Windows, macOS (Apple Silicon), and Linux
Download needs a Pumpkin account with a license. Sign up, pay for the seat, and add your provider keys on first run.